International Conference on Computer and Knowledge Engineering

Home / 14th International Conference on Computer and Knowledge Engineering

Deep Learning-Based Malaysian Sign Language (MSL) Recognition: Exploring the Impact of Color Spaces

Authors :

Ervin Gubin Moung¹ Precilla Fiona Suwek² Maisarah Mohd Sufian³ Valentino Liaw⁴ Ali Farzamnia⁵ Wei Leong Khong⁶

1- Faculty of Computing and Informatics University Malaysia Sabah 2- Faculty of Computing and Informatics University Malaysia Sabah 3- Faculty of Computing and Informatics Universiti Malaysia Sabah 4- Faculty of Computing and Informatics Universiti Malaysia Sabah 5- School of Computing and Engineering University of Huddersfield 6- School of Engineering Monash University Malaysia

Keywords :

sign language،Malaysian Sign Language،color space،ResNet18،Convolutional Neural Network (CNN)

Abstract :

Sign Language is one form of communication for this group of people to communicate with each other. Not only for people with hearing problems but sign language is also useful for people who are mute or have problem speaking. The most used sign language is the American Sign Language (ASL) that is widely used in English speaking countries. In Malaysia, Bahasa Isyarat Malaysia (BIM) or Malaysian Sign Language (MSL) is still new to the community in Malaysia. In this project, a dataset with 5980 images of the signed alphabet is used to train models to recognize what the signs mean. The problem this project aims to address is the limited research and available datasets in the field of Malaysian Sign Language (MSL) recognition using deep learning and various color spaces. Two models that are used are Convolutional Neural Network (CNN) and Residual Network 18 (ResNet18). The images are also converted into different color spaces which are RGB, YCbCr, Grayscale and the combination of RGB and YCbCr. From the results, RGB is the best color space with CNN without any image processing technique - 80% testing accuracy, with Histogram Equalization (HE) - 82.40% testing accuracy, and with Contrast Limited Adaptive HE (CLAHE) - 83.90%. Whereas YCbCr is the best color space when using ResNet18 without any image processing technique - 88% testing accuracy, with HE - 84.40% testing accuracy, and with CLAHE - 88.30%. The precision, recall, and F1-score metrics are also have been used to evaluate the efficacy of the suggested system.